“Machine translation is still broken for most of the world’s languages”: Cohere builds non-reasoning for a reason

Enterprise AI firm Cohere unveiled North Small Translate, a machine translation model capable of handling 50 languages, on a recent week. The open-weight model, available for non-commercial use under CC BY-NC 4.0, is hosted in Cohere's Model Vault for commercial deployment. Cohere touts North Small Translate as part of its sovereign AI strategy, designed for organizations that desire control over their model's deployment and data handling.

The model extends Cohere's multilingual and translation expertise from its Tiny Aya and Command A Translate families. Cohere claims North Small Translate surpasses similarly-sized open-weight models under 1 trillion parameters, as well as API-based translation models in various machine translation dimensions, on average. Cohere co-founder Nick Frosst explains that the model's efficiency stems from its non-reasoning approach, utilizing learned statistical patterns instead of a step-by-step logic process, resulting in fewer tokens used.

Despite machine translation not being perfect for most world languages, Frosst emphasizes that Cohere's approach focuses on smaller, specialized models to be deployed within organizations. The Cohere evaluation using WMT26 benchmarks reveals North Small Translate leading with a score of 83.60, surpassing other models like Qwen 3.5 397B A17B (81.56), GLM 5.2 FP8 (76.50), DeepL NextGen (81.37), Gemma 4 31B (79.46), and Google Translate (68.20).

The model's mixture-of-experts architecture allows for a smaller compute and memory footprint compared to other models. Cohere points out that long documents can become problematic with current solutions, such as a safety manual losing coherence by page ten. Cohere's model addresses this issue, scoring 48.9 on long-context tests.

Additionally, North Small Translate offers enhanced translation performance for complex enterprise tasks, such as structured translations, instruction following, and terminology guides. The model supports 50 languages and includes workflow-focused features like structured translations, instruction following, and terminology guides.

Cohere collaborated with RWS, a language technology and services company, to improve North Small Translate's real-world translation performance. Developers can access the model's weights for non-commercial use, and there's a Hugging Face Space and API available for those without the necessary hardware.